Product funnel detective
By Toolspoke
Reads your PostHog funnel and feature events each week, finds where users drop off, and drafts ranked experiments in Notion.
Needs a Toolspoke account. You choose which project it installs into.
- PostHog + Notion + Slack
- Required integrations
- Runs on a schedule
- Mondays at 9am by default
- Post one summary
- What its last step does
- Some actions can delete
- 1 of 40 actions in its tools
On this page
The short version
- What it is
- A ready-made Toolspoke agent that finds where users drop off in your product and drafts experiments to test.
- What it reads
- Your chosen PostHog funnel, feature events and cohorts, over the last 28 days and the 28 before.
- What it finds
- The steps that lose the most people, the segments where it is worst, and features people stop using.
- What you get
- Up to five ranked experiments in your Notion database, with numbers and replay links, and one Slack summary.
- How often
- Every Monday once you switch it on.
- What it never does
- Change a feature flag, start an experiment, or say a change caused a drop.
- What you need
- A Toolspoke account, a PostHog personal API key, a Notion database and a Slack workspace.
What Product funnel detective does
Product funnel detective is an AI agent that reads your PostHog data every week, finds the step in your funnel where the most people stop, checks which features people adopt, and drafts experiments for your team to consider. You decide what to test.
Most product teams have the data and no time to read it. The funnel chart is on a dashboard, a step dropped six points last month, and nobody noticed until the quarterly review.
Reading it properly takes more than a glance. Is the drop real or a quiet week? Is it everyone, or only mobile users from one campaign? Is a feature quietly being abandoned?
Product funnel detective reads the numbers each week, compares them with the four weeks before, and breaks the worst steps down by segment. It only calls a change real with enough people behind it, and it turns what it finds into experiments, each with the numbers and a few sessions to watch.
In practice, every Monday you have:
- Your funnel, step by step, with conversion and the change on last period
- The segments where the worst steps are worse, such as mobile or one signup source
- Features whose use is falling, or that people try once and leave
- Up to five ranked experiments in Notion, each with a hypothesis and a metric
How it works
Product funnel detective works in four steps each week: it reads the funnel, breaks down the drop-offs, reads feature adoption and drafts the experiments. All of it comes from the agent's written instructions, which you can read and edit after you install it.
Step 1: Read the funnel
From a saved funnel or the events you chose, for two 28 day windows.
Step 2: Break down the drop
By device, browser, signup source and cohort, for the worst steps.
Step 3: Read adoption
Weekly use and repeat use of each feature you track.
Step 4: Draft experiments
Up to five, ranked, with numbers, a hypothesis and replays.
It uses your events, not guesses
Every product names its events differently. At install you choose a saved PostHog funnel or list the events in order, and the features to track. The agent never guesses which event means signed up.
It needs enough people
A step needs at least 200 people at its start in both windows before a change counts, and a segment needs 100 people and a gap of 10 points. Below that it says too few people to tell, rather than proposing an experiment on noise.
It ranks by people lost
Each candidate is scored by how many people the step lost and how much worse it got, and the top five are kept. The numbers are on every experiment, so you can rank it yourself.
It proposes, it does not prove
The data shows where people leave, not why. So each experiment is a hypothesis written as a question, a change to test and the metric that would show it worked. The limits are listed under what it will never do.
What a weekly summary looks like
A run of Product funnel detective drafts experiments in Notion and posts one summary to Slack. The summary leads with the experiments, then what changed in feature use, then what it could not judge.
Northwind
Channels
- general
- product
- growth
- engineering
product
16 Sep to 13 Oct against the 28 days before. Signed up 3,120, created a project 72%, invited a teammate 31% (down 6 points), active in week two 75%.
Experiments proposed (3)
- Half of new projects never get a teammate invited. 1,550 people stopped here, up from 1,290. Worst on mobile: 19% against 34%. Could the invite step come after the first project is saved? 3 replays linked.
- Signups from the partner campaign rarely create a project. 58% against 72% for everyone else, 410 people. Do they land on a page that assumes they know the product?
- The export dialog is opened and closed. Half who open it do not export. Is the format choice unclear? Added to the existing experiment as a comment.
Feature use
- Exports: used by 22% of weekly active people, down from 31%.
- Templates: tried by 40%, used again by 9%.
Too few people to tell
Upgraded to a paid plan: 46 people this window. No change called.
- Experiments proposed
The steps that lost the most people, ranked. Each is a Notion page set to Proposed with the numbers in both windows, the worst segment, a hypothesis, a change to test, a success metric and replay links.
- Feature use
Features whose weekly use fell by a quarter or more, and features people try once and rarely use again.
- Too few people to tell
Steps and segments with too few people to call a change. They are listed so you know they were checked.
When an event goes missing
If an event in your funnel has no data in either window, the summary says which, reports the steps before it and proposes nothing after it. A renamed event looks like a collapse, and the agent will not mistake one for the other.
If PostHog cannot be read, it says so once and stops. If Notion cannot be written, the experiments go in the Slack summary instead.
Who uses Product funnel detective
Product funnel detective suits a product team that tracks events in PostHog and wants someone to read them every week. The examples use made-up names.
A product manager choosing what to test
Hana Sato plans experiments every two weeks. She opens the Proposed view in Notion and starts from the step that lost the most people, with the replays already linked.
A growth lead watching onboarding
Marcus Webb launched a partner campaign. The breakdown shows its signups create a project far less often than everyone else.
A designer looking for friction
Elena Popescu watches the linked sessions for the export dialog before sketching a fix.
A founder with no analyst
Kwame Asante set up PostHog and never had time to read it. The Monday summary is the one place he sees the funnel move.
What Product funnel detective connects to
Product funnel detective connects to three tools: it reads PostHog, drafts and comments in Notion, and posts one summary to Slack. You connect each one yourself, and Toolspoke stores the connection encrypted, so the agent never sees the credential.
PostHog
RequiredRun HogQL queries and read insights, events, flags, cohorts and recordings.
Authentication: Personal API key
Notion
RequiredRead and update pages, databases, and team knowledge.
Authentication: Notion OAuth
Slack
RequiredSearch conversations and send messages to your workspace.
Authentication: Bot token
| Tool | What Product funnel detective does | What it cannot do |
|---|---|---|
| PostHog | Reads saved funnels, runs read-only queries and lists cohorts (get_insight, run_query, list_cohorts). | Change a feature flag if you give it read access. That is the only PostHog action that changes anything, and a read-only grant cannot reach it. Queries are select-only. |
| Notion | Reads your experiments database, drafts new pages and comments on existing ones (notion-fetch, notion-query-data-sources, notion-create-pages, notion-create-comment). | Reach pages you have not shared with the connection. It can update and move pages it can reach, and the agent is told only to draft and comment. |
| Slack | Posts one summary to the channel you chose (send_message). | Read direct messages. It can edit its own messages and schedule messages, and the agent is told to do neither. |
That the agent cannot change a flag is a lock, if you install PostHog with read access. That it only drafts and comments in Notion, and posts one message in Slack, is an instruction to the agent, not a lock. If you want a lock, a project can switch individual actions off. The access control guide explains how.
PostHog holds data about your users. The agent reports people as counts and never copies an email address, a name or a person property into Notion or Slack. A replay link is how someone on your team sees an individual session, inside PostHog.
What Product funnel detective will never do
Product funnel detective is told never to change a feature flag, never to start an experiment, and never to say a change caused a drop. It finds and proposes, and a person decides what to test.
- Change, create or archive a feature flag, start an experiment, or edit an insight or dashboard.
- Say a change caused a drop, or that an experiment will work.
- Call a difference real on fewer people than the thresholds above.
- Copy an email address, a name or any person property into Notion or Slack.
- Change an experiment's status, or take one out of Proposed.
Why it will not touch your product
A feature flag changes what real users see the moment it flips. The agent reads PostHog and nothing else, and with a read-only key it cannot reach the one action that changes a flag.
Why it will not claim a cause
A funnel shows where people leave, not why. A drop might be a design problem, a new campaign bringing different people, or a renamed event. The agent writes what it saw and a question to test.
Why it will not call noise a trend
Small numbers swing. Below 200 people at a step, or 100 in a segment, it says too few people to tell and proposes nothing.
If something goes wrong
If PostHog cannot be read, post that once and stop. If an event in the funnel has no data in either window, say which, report the steps before it, and propose nothing after it: the event may have been renamed. If Notion cannot be written, put the experiments in the Slack summary so the work is not lost. If the database has no Proposed status, create nothing and say so.
These are written into Product funnel detective's instructions. They guide the agent, but they are not a permission. What it is allowed to touch is set by the access you grant.
Product funnel detective reads the numbers and drafts experiments. What to test, what to ship and every change your users see stays with your team.
When Product funnel detective runs
Product funnel detective runs every Monday at 9 AM, but only after you switch its schedule on. It arrives switched off, so nothing runs and nothing is charged until you decide.
- Each run compares the last 28 days with the 28 days before.
- It drafts up to five experiments, and comments on ones that already exist.
- It posts one summary to the Slack channel you chose.
- If an event has no data, it says which and proposes nothing after it.
Each run uses model credit from your workspace balance, so it makes one run a week. The queries run in PostHog and count against your PostHog plan in the usual way. The Toolspoke tool calls themselves are free.
Your copy is an ordinary agent in your project. You can pause its schedule there, run it once on demand, change the day and edit its instructions. The agent's page lists recent runs and their cost.
How to set up Product funnel detective
To start finding drop-offs, add Product funnel detective to a Toolspoke project, connect PostHog, Notion and Slack, name your funnel and switch on the schedule. It takes a few minutes, and nothing runs until the last step.
Step 1: Create a Toolspoke account or sign in
Choose Use this agent. If you are new, you will be asked to sign up first, and you come straight back to this agent afterwards.
Step 2: Add it to a project
Pick the project it should live in. You get your own editable copy, and you can see the three tools it brings before anything connects.
Step 3: Connect PostHog with read access
Create a personal API key in PostHog with read scopes and paste it, with your PostHog host. Read access means it cannot change a feature flag.
Step 4: Name your funnel and features
Choose a saved funnel insight, or list your funnel's events in order, and the feature events to track. Use your event names exactly as PostHog shows them.
Step 5: Connect Notion and Slack
Share your experiments database with Notion, with a Status property that has a Proposed option. Sign in to Slack and invite the bot to the summary channel.
Step 6: Switch on the schedule
The schedule arrives switched off. Run it once, check the funnel numbers against PostHog, then switch it on.
Make Product funnel detective fit your product
Product funnel detective's instructions are plain English and you can edit them after you install it. Installing copies the agent, so your edits stay with your copy. Open the agent in Toolspoke and tell the builder what to change, and it rewrites the instructions.
Paste any sentence below into the builder chat, or say the same thing in your own words.
Also compare conversion by the person property plan, and name a plan when it converts 10 points worse.
Call a change real at 80 people per step instead of 200, and say the numbers are small.
Also watch the funnel viewed_pricing, started_checkout, paid, and report it separately.
Put the experiments in the Slack summary only, and create nothing in Notion.
Run every second Monday at 9 AM.
This changes when the agent wakes up, so the builder edits its schedule, not the instructions.
None of these edits needs more access or touches the Never rules. They change what Product funnel detective is told, not what the connections allow.
How it compares to other options
Product funnel detective beats the alternatives when you have PostHog data and nobody reading it every week, and loses to them when you need deep, one-off analysis or a judgement call about your users.
| Compared on | Product funnel detective | A PostHog dashboard | A product analyst | A quarterly review |
|---|---|---|---|---|
| Reads the funnel every week | Yes | If someone looks | Yes | No |
| Breaks drops down by segment | Yes, the worst steps | If you build it | Yes, in depth | Sometimes |
| Ignores changes too small to trust | Yes, by threshold | No | Yes | Varies |
| Turns findings into experiments | Yes, drafts in Notion | No | Yes | Sometimes |
| Knows why users leave | No, it proposes tests | No | With research | Opinions |
| Cost | A run a week | Included | A salary | A day of meetings |
A product analyst is the better choice when you need to understand why, with interviews, deep segmentation and judgement about your market. Product funnel detective is for the team that has PostHog and no one reading it every week.
A PostHog dashboard is where the numbers live, and the agent reads the same data. The difference is that it reads it every Monday, compares it with last month, and writes down what to test.
Frequently asked questions
Can Product funnel detective change my product?
No. It reads PostHog. The only PostHog action that changes anything is updating a feature flag, and with a read-only key the connection cannot reach it.
How does it know which events matter?
You tell it at install: a saved funnel, or your events in order, and the features to track. It never guesses which event means what.
Will it tell me why users drop off?
No. The data shows where people leave, not why. It proposes a hypothesis and a test, and links a few sessions to watch. The test tells you why.
What if we do not have many users?
Below 200 people at a step it says too few people to tell and proposes nothing on that step. You can lower the threshold by editing its instructions, and it will say the numbers are small.
Does it put user data in Notion or Slack?
No. It reports people as counts and never copies an email, a name or a person property. Replay links open inside PostHog, for people on your team who can already see them.
Does it work with Amplitude or Mixpanel?
Not as shipped, because it reads PostHog. Toolspoke has Amplitude and Mixpanel toolkits, so you can ask the agent builder for another version. It needs its own instructions.
Is Product funnel detective free?
It has no price of its own, and every Toolspoke plan includes agents. Tool calls cost nothing, but each run draws on your usage balance. Pricing has the details.
Find the step where users leave
Connect PostHog, name your funnel and switch on the schedule. Your first set of experiments lands on Monday.